GLM-5.2 and IndexShare for Long-Context Sparse Attention

Z.ai released GLM-5.2, an open-weight model that the author calls the best open-weight model available. The model introduces IndexShare, a cross-layer reuse trick for DeepSeek Sparse Attention that reduces the cost of 1M-token inference by running the sparse-attention indexer only once every four layers. GLM-5.2 scored 68.8 on the Artificial Analysis Coding Index, outperforming Claude Opus 4.8 by more than 10 points.

GLM-5.2 and IndexShare for Long-Context Sparse Attention GLM-5.2 https://huggingface.co/zai-org/GLM-5.2 is a recent open-weight model release from Z.ai. My first impression is that it is the best open-weight model today. As usual for fresh releases, I would treat the release-time leaderboard position as date-sensitive. Architecture-wise, it builds on the earlier GLM-5 and GLM-5.1 architecture. In particular, it reuses Multi-head Latent Attention /llm-architecture-gallery/mla/ and DeepSeek Sparse Attention /llm-architecture-gallery/deepseek-sparse-attention/ , the DSA mechanism from DeepSeek V3.2 that I covered in the DeepSeek V3 to V3.2 article https://magazine.sebastianraschka.com/p/technical-deepseek . What’s new is IndexShare. This is a cross-layer reuse trick for DSA. Instead of recomputing the sparse-attention top-k indexer in every layer, GLM-5.2 runs the full indexer only once every four layers. The following layers then reuse the selected token indices. This keeps the same DSA idea but makes 1M-token inference cheaper. The attention pattern is still adaptive, but the model spends less work repeatedly deciding which earlier tokens to attend to. The local GLM-5.2 architecture card /llm-architecture-gallery/ card-glm-5-2 has the current summary, config links, and benchmark /glossary/ benchmark references. By the way, in the Artificial Analysis Coding Index https://artificialanalysis.ai/ snapshot below, GLM-5.2 scores 68.8 versus 56.7 for Claude Opus 4.8 max , more than 10 points higher on coding benchmarks. That’s super impressive for a fresh open-weight model. Source: lightly edited website version of my Substack note https://substack.com/@rasbt/note/c-278515750 . Read Next Inkling: A New Open-Weight 975B MoE with a Few Surprises Short note on Thinking Machines Lab's 975B Inkling open-weight model, its benchmark profile, sparse MoE design, short convolutions, embedding RMSNorm, and /blog/2026/inkling-architecture-benchmark-notes.html 200,000 Subscribers Short note celebrating Ahead of AI reaching 200,000 subscribers. /blog/2026/ahead-of-ai-reached-200000-subscribers.html GPT 5.6 Has 72 Possible Configurations. What's A Good Default? Short note on how GPT 5.6 model and effort choices map onto training-time and inference-time scaling, producing 72 configurations. /blog/2026/gpt-5-6-configurations.html